| name | rijksmuseum-mcp |
| version | 0.23.2 |
| last_updated | "2026-03-29T00:00:00.000Z" |
| description | Research workflows for the Rijksmuseum MCP+ server. Use this skill whenever the user wants to search the Rijksmuseum collection, look up artwork details, explore iconographic themes, inspect high-resolution images, trace provenance, browse curated sets, or answer any art-historical question that could be addressed through the Rijksmuseum's holdings of ~832,000 artworks.
|
Rijksmuseum MCP+ Research Skill
Core Mental Model
The server is not a 'one and done' search engine โ it is a resource for actively exploring and researching the collection of the Rijksmuseum. Many meaningful questions will require several chained tool calls. The canonical sequence is:
DISCOVER โ QUANTIFY โ RETRIEVE โ INSPECT
Avoid jumping directly to RETRIEVE without a DISCOVER or QUANTIFY step unless the query is simple and straightforward (e.g. the object number is already known). For QUANTIFY, use collection_stats for distributions and cross-domain breakdowns, compact: true for simple counts.
Tool Selection Guide
| Question type | Start here |
|---|
| "Find works about X" โ clear vocabulary concept | search_artwork with subject |
| "Find works about X" โ interpretive / atmospheric | semantic_search |
| "Find works depicting [iconographic scene]" | lookup_iconclass โ search_artwork with iconclass |
| "How many works match X?" | collection_stats for counts/distributions; search_artwork with compact: true only when you need object numbers |
| "Distribution of X across the collection" | collection_stats with dimension |
| "Top N creators / depicted persons / places" | collection_stats with dimension + topN |
| "Sales by decade" / time series | collection_stats with dimension: "provenanceDecade" |
| "How many artworks have LLM-mediated interpretations?" | collection_stats with dimension: "categoryMethod" |
| "Of artworks with provenance, how many are paintings?" | collection_stats with dimension: "type" + hasProvenance: true |
| "What does the Rijksmuseum say about this work?" | get_artwork_details |
| "What scholarship exists on this work?" | get_artwork_bibliography |
| "Examine this image closely / read this inscription" | inspect_artwork_image |
| "Show me a curated group of works on X" | list_curated_sets โ browse_set |
| "Has anything changed / been acquired recently?" | get_recent_changes |
| "Who owned this work / trace its ownership chain" | search_provenance with objectNumber |
| "Which works passed through collector X?" | search_provenance with party |
| "Find confiscations / sales / transfers in city Y" | search_provenance with transferType, location |
| "How long did family X hold their collection?" | search_provenance with layer: "periods", ownerName |
Choosing between search_artwork(provenance=...) and search_provenance:
search_artwork with provenance is a keyword search over raw provenance text โ fast for counting, faceting, and cross-tabulation (e.g. "how many Mannheimer works are porcelain?"). search_provenance returns structured, parsed chains with dates, prices, transfer types, and ownership periods โ use it when you need to reason about the sequence of ownership, filter by event type, or rank by price or duration. For the last link in the chain โ how the Rijksmuseum acquired a work โ also check creditLine, which covers ~358K artworks (vs ~48K with parsed provenance) and often names donors or funds absent from the provenance chain (e.g. "Drucker-Fraser", "Vereniging Rembrandt").
Critical Parameter Distinctions
productionPlace vs depictedPlace
These are semantically opposite and getting them wrong produces systematically
wrong results.
productionPlace: where the object was made โ "Amsterdam", "Delft"depictedPlace: what place the object shows โ "Batavia", "Nagasaki"
A Dutch painting of Batavia was made in Amsterdam. Use both together to
distinguish objects from Asia versus European images of Asia.
subject vs iconclass
subject: morphological stemming against ~108K terms (almost all in English); best first pass; try natural phrases ("winter landscape", "vanitas", "civic guard")iconclass: the Iconclass classification system โ a universal art-historical taxonomy of ~40,675 notations across 13 languages, with hierarchical browsing and pre-computed Rijksmuseum artwork counts. Three search modes via lookup_iconclass:
- keyword (
query): FTS5 across labels and keywords in all 13 languages (exact word match)
- semantic (
semanticQuery): embedding-based concept search โ finds notations by meaning, not just words (e.g. "domestic animals" finds dogs, cats, horses)
- browse (
notation): navigate the hierarchy โ view any notation's path, children, cross-references, and keywords
Each result includes a rijksCount (how many Rijksmuseum artworks carry that notation), so you can gauge usefulness before searching. Pass the notation code to search_artwork's iconclass parameter for precise filtering.
Decision rule: start with subject โ it's faster and handles most queries well. Switch to iconclass when:
- you need to distinguish closely related scenes (Crucifixion vs Deposition, Annunciation vs Visitation)
subject returns too broad a result set and you need hierarchical precision- you want to explore a conceptual neighbourhood (browse children/siblings of a notation)
- the query is in a non-English language (Iconclass covers 13 languages; subject vocab is almost entirely English)
- you want to search by meaning rather than keywords (semantic mode: "religious suffering", "festive gathering")
dateMatch โ controlling how date ranges are binned
Most artworks have date ranges (e.g. "1630โ1640"), not exact years. dateMatch controls how these ranges interact with creationDate wildcards and collection_stats decade/century dimensions:
| Mode | Behaviour | Use when |
|---|
"overlaps" (default) | Artwork appears in every bin its range touches | Inclusive discovery โ "show me everything that could be from the 1630s" |
"within" | Artwork appears only if its entire range falls within the bin | Exclusive bins, but drops ~43% of the collection (broadly-dated objects) |
"midpoint" | Each artwork counted in exactly one bin (midpoint of its date range) | Statistical comparisons, charts, decade-by-decade counts โ no double-counting, no data loss |
Rule of thumb: use midpoint whenever you're comparing counts across bins (decade trends, century breakdowns, gender ratios by period). Use the default overlaps for discovery queries where inclusiveness matters more than precision.
Available on search_artwork, semantic_search, and implicitly via collection_stats creationDateFrom/creationDateTo filters.
attributionQualifier + creator โ structural limitation
The parameter description suggests combining attributionQualifier with creator to filter by source artist (e.g. "find all followers of Rembrandt"). This combination never works โ the tool schema documentation is incorrect. Verified across all qualifier types (follower of, workshop of, circle of, attributed to): the structured creator field is always Unknown [painter] or anonymous, regardless of qualifier. The source artist's name appears only in the composite display string and is not exposed as a searchable entity field.
The correct strategy by qualifier type:
| Goal | Working approach |
|---|
| Works in the manner/style of artist X | aboutActor: "X" + type: "painting" (or other type) |
| All "follower of" works in the collection | attributionQualifier: "follower of" alone (returns ~111 works) |
| Sub-filter those by source artist | Not possible via parameters โ requires fetching and inspecting individual records |
Additional complication โ "manner of" is not a valid qualifier value. The Rijksmuseum uses "manner of" as a distinct attribution category, but "manner of" is not in the attributionQualifier controlled vocabulary and returns zero results. Works attributed in this way are retrievable only via aboutActor.
Canonical name form matters. The Rijksmuseum catalogue uses historical Dutch/Latin spellings for some artists. Bosch is catalogued as "Jheronimus Bosch", not "Hieronymus Bosch". Always check get_artwork_details on a known work to confirm the canonical form before filtering.
creator vs aboutActor vs depictedPerson
creator: who made it โ uses canonical name forms ("Rembrandt van Rijn")depictedPerson: who is shown in it โ matches 210K name variants including historical formsaboutActor: broader cross-field search; more tolerant of name variants and cross-language forms ("Louis XIV" finds "Lodewijk XIV")
search_provenance: two query layers
search_provenance has two data layers that answer fundamentally different questions.
layer: "events" (default): individual transactions โ each with a date, location, price, parties (with roles and positions), and transfer type. Think of it as a ledger of what happened. Events-only parameters: transferType, excludeTransferType, hasPrice, currency, hasGap, relatedTo.layer: "periods": interpreted ownership spans โ who held the work, how they acquired it, and for how long. Think of it as a timeline of who owned what. Periods-only parameters: ownerName, acquisitionMethod, minDuration, maxDuration, sortBy: "duration".
Shared parameters work on both layers: party, location, creator, dateFrom/dateTo, objectNumber, categoryMethod, positionMethod, sortBy, offset, facets.
dateFrom/dateTo semantics differ by layer:
- Events: filters on the event's
date_year โ "something happened between these years"
- Periods:
dateFrom filters on begin_year, dateTo on end_year โ "ownership that started after X AND ended before Y"
The periods interpretation is much more restrictive โ dateFrom=1933, dateTo=1945 on periods misses any ownership that started before 1933 or extended past 1945. For date-range queries (especially wartime provenance), prefer the events layer.
Anti-join pattern (transferType + excludeTransferType): artwork-level set difference. transferType: "confiscation", excludeTransferType: "restitution" returns artworks that were confiscated but never restituted. Note: items recuperated (recovered by Allied forces) are not the same as items restituted (formally returned to original owners) โ they will appear in anti-join results.
Filter requirement: both layers reject bare queries. At least one content filter is required. If you need a collection-wide ranking, use a broad filter such as dateFrom: 1400 as a catch-all.
For the full provenance data model โ AAM text format, transfer type vocabulary, party roles and positions, date/currency representations, and tested query patterns โ see references/provenance-patterns.md.
Full-text filters vs vocabulary filters (ranking matters)
Results rank by BM25 text relevance when any of these are active:
title, description, inscription, provenance, creditLine,
curatorialNarrative.
Results rank by importance (image availability + curatorial attention +
metadata richness) when only vocabulary or column filters are active.
Practical rule: if you want the most significant works to surface first,
use vocabulary filters alone. If you want keyword relevance, activate a text
filter.
Modifier Parameters (cannot stand alone)
These narrow results but require at least one other content filter:
| Modifier | Notes |
|---|
imageAvailable: true | Always pair with a content filter |
hasProvenance: true | Restricts to ~48K artworks with parsed provenance. Pair with type, creator, etc. for cross-domain queries. |
creatorGender: "female" | Pair with type, subject, or creationDate |
creatorBornAfter / creatorBornBefore | Pair with any content filter |
expandPlaceHierarchy: true | Expands place filters 3 levels deep; pair with productionPlace etc. |
Key Workflows
1. Scope Before You Browse
For any counting, distributional, or comparative question, start with collection_stats โ it answers in one call what would otherwise require multiple compact: true loops.
# Example: Rembrandt's output across media ๏ฟฝ๏ฟฝ๏ฟฝ one call
collection_stats(dimension="type", creator="Rembrandt van Rijn")
# โ painting 314 (38.2%), print 289 (35.2%), drawing 218 (26.5%)
# Cross-domain: what types of artworks have provenance events in Amsterdam?
collection_stats(dimension="type", hasProvenance=true, location="Amsterdam")
# Top 30 depicted persons
collection_stats(dimension="depictedPerson", topN=30)
# Cross-filter with Iconclass or curated sets
collection_stats(dimension="type", iconclass="73D82")
# โ what types of artworks depict the Crucifixion?
collection_stats(dimension="creator", collectionSet="Japanese prints", topN=20)
# โ top 20 creators in the Japanese prints set
# Physical size distributions (binned by binWidth, in cm)
collection_stats(dimension="height", type="painting", binWidth=20)
# โ height distribution of paintings in 20cm bins
Use compact: true on search_artwork only when you need the actual object numbers (e.g. to feed into get_artwork_details), not just the count or distribution. For pure counting, collection_stats is always more efficient.
2. Iconclass Two-Step
Never pass iconographic concepts as free text to search_artwork(iconclass=...) โ
it expects exact notation codes. Always discover valid codes first via lookup_iconclass.
Step 1: discover notation codes (three modes available)
lookup_iconclass(query="Crucifixion") # keyword
lookup_iconclass(semanticQuery="religious suffering", onlyWithArtworks=True) # semantic
lookup_iconclass(notation="73D8") # browse hierarchy
# โ shows children: 73D81 (Christ before Pilate), 73D82 (Crucifixion), โฆ
Step 2: search with the discovered code
search_artwork(iconclass=["73D82"])
search_artwork(iconclass=["73D82", "25F33(DOVE)"]) # AND-combines multiple codes
Check rijksCount in lookup_iconclass results before searching โ a code with rijksCount: 0 will return nothing.
3. Century / Decade Wildcards
For longitudinal analysis, prefer collection_stats with decade or century binning. Use dateMatch: "midpoint" for statistical comparisons โ the default overlaps mode double-counts broadly-dated objects across bins (see Critical Parameter Distinctions above):
collection_stats(dimension="decade", technique="etching", creationDateFrom=1500, creationDateTo=1800)
# โ decade-by-decade etching counts in one call (default overlaps โ inclusive but inflated)
collection_stats(dimension="decade", technique="etching", creationDateFrom=1500, creationDateTo=1800, binWidth=50)
# โ half-century bins
For queries that need individual object numbers (not just counts), use creationDate wildcards with search_artwork:
search_artwork(technique="etching", creationDate="16*", dateMatch="midpoint")
search_artwork(technique="etching", creationDate="17*", dateMatch="midpoint")
# dateMatch="midpoint" ensures each artwork is counted in exactly one century
Supported patterns: "1642" (exact year), "164*" (decade), "16*" (century), "1*" (millennium).
4. Semantic Search + Structured Verification
Use semantic_search for concepts with no vocabulary term or Iconclass code โ
atmosphere, emotion, cultural interpretation. Always follow up with a
structured search_artwork to test whether the same works are reachable
through controlled vocabulary.
semantic_search(query="loneliness and isolation in a vast empty space", type="painting")
# Review source text for each result โ ask why it appeared
# Then: search_artwork(subject="landscape", curatorialNarrative="solitary")
# Compare overlap: works only reachable via semantic search are the interesting gap cases
Pre-filtering: semantic_search accepts all vocabulary filters from search_artwork as pre-filters โ subject, iconclass, depictedPerson, depictedPlace, productionPlace, collectionSet, aboutActor, imageAvailable, dateMatch. These narrow the candidate set before semantic ranking, combining structured precision with conceptual search:
semantic_search(query="vanitas symbolism", subject="skull", type="painting")
# โ paintings with skull subjects, ranked by how well they match "vanitas symbolism"
Language note: English queries yield slightly higher precision against the
bilingual catalogue even though the embedding model is multilingual. If a Dutch
or German query returns unexpected results, reformulate in English.
5. Image Inspection and Overlay Placement
Basic inspection + zoom is now a single step. inspect_artwork_image automatically navigates the open viewer to the inspected region (navigateViewer defaults to true). No separate navigate_viewer call is needed for basic zoom โ the viewer stays in sync with your analysis.
# Single call: inspect a region AND zoom the viewer there
inspect_artwork_image(objectNumber="SK-C-5", region="pct:70,60,20,20")
# โ base64 image for AI analysis + viewer auto-zooms to the same region
For overlays, use the inspect โ overlay two-step with relativeTo:
# Step 1: inspect the area
inspect_artwork_image(objectNumber="SK-C-5", region="pct:70,60,20,20")
# โ found a detail at roughly the center of this crop
# Step 2: place overlay using crop-local coordinates โ the server projects to full-image space
navigate_viewer(viewUUID=..., commands=[{
action: "add_overlay",
region: "pct:30,30,40,40", # coordinates within the crop
relativeTo: "pct:70,60,20,20", # the crop region from Step 1
label: "Detail"
}])
The relativeTo parameter eliminates manual coordinate conversion โ specify where the feature is within the crop, and the server handles the projection. Both region and relativeTo must use pct: format.
Use region="full" for an initial composition overview before cropping to
details. inspect_artwork_image can surface content absent from structured
metadata โ unsigned Japanese prints often have readable artist signatures,
publisher seals, and poem cartouches that the catalogue has not transcribed.
6. Provenance and Acquisition Research
Provenance research moves through three levels of detail:
-
Scope and profile โ search_artwork with provenance (keyword) or
creditLine for fast counts and facets. creditLine covers ~358K artworks
(vs ~48K with parsed provenance) and captures the last link: how the museum
acquired the work. Combine with productionPlace + expandPlaceHierarchy
for geographic cross-tabulation.
-
Structured chain analysis โ search_provenance for parsed chains with
dates, prices, transfer types, and ownership periods. Add facets: true
for quick distributional context alongside chain results.
-
Single-object deep dive โ search_provenance(objectNumber=...) for the
full chain, then get_artwork_details for narrative text and
get_artwork_bibliography for scholarly references. Use bibliographyCount
in get_artwork_details to triage before committing to full citation fetches.
Cross-domain queries: hasProvenance: true on search_artwork or
collection_stats bridges the two systems โ e.g.
collection_stats(dimension="type", hasProvenance=true).
Enrichment transparency: categoryMethod and positionMethod are
queryable input filters on search_provenance, not just output fields. When
results contain LLM-enriched records, the response includes a review URL.
Always show this URL to the user.
Pagination: when totalResults exceeds 50, paginate with offset
(increments of 50 until offset โฅ totalResults).
For the complete data model (AAM text format, transfer type vocabulary, party
roles, date and currency representations), tested query patterns (collector
profiling, wartime provenance, price history, acquisition channels, decade-level
time series), and enrichment methodology, see references/provenance-patterns.md.
7. SourceโCopy Navigation
relatedObjects in get_artwork_details provides direct links to associated
works. Use it instead of a second search when tracing reproductive prints back
to painted sources.
search_artwork(productionRole="after painting by", creator="Rembrandt van Rijn")
# โ get_artwork_details on a result
# โ relatedObjects includes the source painting's Linked Art URI
# โ get_artwork_details(uri="https://id.rijksmuseum.nl/200666460") for full source metadata
# โ get_artwork_image on both for side-by-side comparison
8. Collection Depth Assessment
For grant applications or scoping a research site:
collection_stats(dimension="creator", type="print", productionPlace="Japan", topN=20)
# โ top 20 print artists from Japan + total count, in one call
collection_stats(dimension="decade", type="print", productionPlace="Japan")
# โ temporal distribution
list_curated_sets(query="Japan") # curatorial groupings
browse_set(setId="...") # range of artists/dates
search_artwork(productionPlace="Japan", type="print", maxResults=10) # sample works for closer inspection
9. Gender and Demographic Analysis
creatorGender is a modifier โ always combine it with at least one content
filter. Use century wildcards to reveal structural patterns over time.
# Count women painters per century โ use dateMatch="midpoint" to avoid double-counting
search_artwork(creatorGender="female", type="painting", creationDate="17*", dateMatch="midpoint", compact=true)
# Repeat for "18*", "19*"; then drop compact=true to identify leading artists per century
# Or: one-call gender breakdown per century
collection_stats(dimension="creatorGender", type="painting", creationDateFrom=1700, creationDateTo=1799)
# โ male 4,521 (96.3%), female 48 (1.0%), ...
Known Limitations and Fallbacks
| Issue | Workaround |
|---|
navigate_viewer WebSocket disconnection | Use inspect_artwork_image as the reliable fallback for region analysis |
No subject results in English | Try the Dutch term โ vocabulary is bilingual ("fotograaf" not "photographer") |
semantic_search skews toward prints/drawings | Filter with type: "painting" โ prints and drawings outnumber paintings ~77:1 |
bibliographyCount is high but citations needed | Use get_artwork_bibliography(full=false) to get structured metadata without full text |
| Inscription field empty in catalogue | Use inspect_artwork_image โ AI vision can often read text directly from the image. Try quality: "gray" for better contrast on faded inscriptions or signatures. |
facets on search_artwork | 11 dimensions: type, material, technique, century, creatorGender, rights, imageAvailable, creator, depictedPerson, depictedPlace, productionPlace. Configure with facetLimit (1โ50, default 5). Pass facets=true for all or facets=["creator","type"] for specific ones. All entries include percentage. |
facets on search_provenance | 5 dimensions: transferType, decade, location, transferCategory, partyPosition. Set facets=true. Percentages included. |
| Collection-wide distributions | Use collection_stats instead of compact=true counting loops โ 21 dimensions across artwork and provenance domains with cross-domain filters (including height/width for physical size distributions, binned by binWidth). One call replaces N iterations. |
search_provenance data model details | Batch prices, unsold lots, historical currencies, gap flags, parse methods, party coverage, 0-year durations, and sortBy: "eventCount" unreliability โ see references/provenance-patterns.md |
attributionQualifier + creator | Does not work โ use aboutActor instead. See Critical Parameter Distinctions above. |
attributionQualifier: "manner of" | Not a valid value โ returns zero results. Use aboutActor. |
| Canonical artist name forms | Some artists use historical spellings (e.g. "Jheronimus Bosch"). If a known artist returns no results, check the canonical form via get_artwork_details on a known work. |
Output Conventions
- Always report
objectNumber (e.g. SK-C-5) linked to the Rijksmuseum's page for that artwork when surfacing works โ it is the stable identifier across all tools
- Use
persistentId (handle.net URI) for any citation or bibliographic reference
- When presenting multiple works, lead with the most significant first (the default importance ranking handles this for vocabulary queries)
- For image inspection findings, distinguish explicitly between what the structured metadata says and what the AI reads directly from the image โ the distinction matters for research rigour
